Mistral Unveils 1-Trillion-Parameter ‘Large 4,’ Europe’s Strongest Open-Weight Model
A day after Reflection AI launched its open-weight model Beam, saying it “advances the Western open-weight frontier,” a bigger model has arrived from Europe. Open-weight large language models have effectively been a race led by Chinese companies, with Western newcomers like Reflection AI and Thinking Machines Lab only recently starting to close the gap. Now Mistral AI, already Europe’s leading sovereign-AI company, has joined in — and the Western open-weight bench just got noticeably deeper.

Mistral said it has started a public preview of its new flagship model, Mistral Large 4, internally nicknamed “le Chonk.” It’s a sparse mixture-of-experts model with 1 trillion total parameters, 49 billion of which are active at inference, and it’s natively multimodal from the ground up, handling text and images together. Mistral says the model performs on par with the world’s strongest open-source models while significantly outperforming any other open-weight model built in the US or Europe. Weights are set to ship later this month; until then, the company is red-teaming the same model — with reduced moderation and expanded cyber capabilities — with cybersecurity experts, vetted partners, and government authorities. Its predecessor, Large 3, shipped under an Apache 2.0 license, but Large 4’s license terms haven’t been announced yet.
A Cybersecurity Model That Doesn’t Refuse
Cybersecurity is where Mistral is pushing hardest. On the Artificial Analysis Cyber Index, which measures how well a model finds and fixes real software security flaws, Large 4 ranks among the top five models globally, and it scored 82% — the highest of any model — on a test that asks a model to reproduce a real vulnerability in open-source software and then patch it. It also solved 93% of Cybench, a set of 40 exercises drawn from security competitions, one of the best scores reported for an open-weight model.
There’s a paradox behind that high score. Leading closed models like Claude Opus 5.5 and GPT-6 Astra score near zero on the same tests, because they simply refuse to do the work — and defending software often starts with proving a flaw is real, exactly the kind of task closed-model safety filters tend to block. Mistral argues that as attackers increasingly jailbreak those same closed models to support offensive cyber activity, defenders need systems that can match that capability without the same refusals. At the same time, Mistral says Large 4’s average refusal rate on prompts with cyber-abuse potential is actually higher than other open-weight models — an attempt, it says, to balance defensive capability against misuse.
On coding and agentic tasks, it also posts strong scores — 61.7% on DeepSWE 1.1, 28.3% on Terminal-Bench 4.0 — and beats Kimi K3 and DeepSeek V4 Pro on AutomationBench, a business-workflow-automation benchmark. In a blind human evaluation where professional annotators rated coding output with model identities hidden, Large 4 ranked second of five models, behind only Claude Opus 5 and ahead of Kimi K3, GLM-5.3, and GLM-5.2.
Trained on 3,800 GPUs in Mistral’s Own European Data Centers
Large 4 was trained end-to-end on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers, and the current preview runs on the same infrastructure. The company said a significant share of its training data was multilingual, spanning more than 160 languages including every official EU language, and that its European service runs end-to-end under Mistral’s own control, under European law, independent of other digital service providers.
The model is also the first product to come out of the €3 billion Series D that Samsung Electronics led in September, which pushed Mistral’s valuation above €21 billion — the largest equity round ever raised by a European tech company. Mistral says it’s using that capital to aggressively scale compute capacity at its own European data centers, and that the training, customization, and reinforcement-learning environment behind Large 4 is the same one it sells to enterprise customers as Mistral Forge. More than 125 global enterprises, including Airbus, ASML, and HSBC, are already running mission-critical AI work with Mistral, the company said.
Competitors: A Third Camp Joins a Three-Way Race
Mistral competes across three broad camps. Closed-model labs — OpenAI, Anthropic, Google — lead on raw performance but don’t release weights and tend to lock customers into their own ecosystems. Open-weight players include Meta’s Llama, Google’s Gemma, and Chinese companies like DeepSeek, Zhipu AI’s GLM, Moonshot AI’s Kimi, and Alibaba’s Qwen — all of which release weights but, unlike Mistral, don’t also build their own infrastructure to offer “full-stack sovereignty.” That open-weight camp has recently gained Western entrants like Reflection AI and Thinking Machines Lab, cracking what had been a heavily China-dominated field.
In European sovereign AI specifically, Cohere is Mistral’s most direct rival. Cohere grew by acquiring Germany’s Aleph Alpha in April, and the combined company’s PhariaAI platform offers a sovereign stack that stays entirely within German jurisdiction, aimed at public-sector and regulated-industry customers. Mistral’s pitch is that few rivals combine a European home base, open-weight models, self-built infrastructure, and now, with Large 4, top-tier benchmark performance all at once.
MORE FROM THE POST
- Reflection AI Launches Open-Weight Model ‘Beam,’ Pushing the Western Open Frontier
- Thinking Machines Lab Unveils Inkling 2 Small as It Reportedly Nears a $5B Round
- Meta Releases Open-Weight ‘Muse Spark Open,’ Optimized for Apple Mac Studio
- NVIDIA Plans Investments in Korean AI Startups Motif, Upstage and RLWRLD
- Meta Unveils Muse Wonder and Wonder Full Agent Models
Share
Most Read
- 1
- 2
- 3
- 4
- 5


Leave a Reply